Detection Settings

Configure thresholds and optimize detection performance

Confidence Threshold

The confidence threshold filters detections by their probability score. Only detections above this threshold are shown.

Low (0.15 - 0.25)

More detections, including uncertain ones. Good for initial scanning.

Medium (0.25 - 0.50)

Balanced detection. Recommended for most use cases.

High (0.50 - 0.80)

Fewer but more confident detections. May miss some assets.

Recommendation: Start with 0.25 and adjust based on your results. Lower the threshold if you're missing assets, raise it if you're getting too many false positives.

IoU Threshold (NMS)

Intersection over Union (IoU) threshold controls Non-Maximum Suppression (NMS). It determines when overlapping detections should be merged.

How it works:

If two boxes overlap more than the IoU threshold, the lower-confidence one is suppressed
Lower IoU (0.3): More aggressive suppression, fewer overlapping boxes
Higher IoU (0.7): Less suppression, allows more overlapping detections

Note: Most ONNX models have NMS built-in. This setting primarily affects TFJS models or custom post-processing.

Detection Frame Rate

Controls how often detection runs during video playback. Higher rates detect more assets but use more CPU/GPU resources.

Setting	FPS	Use Case
Economy	2-5	Battery saving, older devices
Balanced	10-15	Recommended for most users
Performance	20-30	Real-time detection, powerful GPUs

Backend Selection

RoadAsset automatically selects the best available backend for inference:

WebGPU Fastest

Hardware-accelerated GPU inference. Requires Chrome 113+ or Edge 113+.

WebGL

GPU acceleration via WebGL. Good compatibility, slightly slower than WebGPU.

WASM (WebAssembly)

CPU-based fallback. Works everywhere but slower. Uses SIMD if available.

Performance Tips

Use Chrome or Edge for WebGPU acceleration
Close other GPU-intensive tabs during detection
Use ONNX models for best inference speed
Lower frame rate on battery power
Cached models load instantly after first download